Stratigraphic segmentation using convolutional neural networks
The Discovery of new possibles reserves is an critical activities for the oil and gas industry. The most used methods to understand the sub-surface are based on seismic surveys. however, the process of the interpretation of these surveys are very expensive and due to the volume of data it overload the human capabilities. On the other hand, deep learning techniques have been increasingly applied in several areas of science to help in tasks that were considered human-centered, such as image classification and language translation, among others. We propose a machine learning methodology to classify seismic data at the pixel level, producing an interpretation mask suggestion. Our methodology comprises three main parts: model selection, dataset preparation, and training. We also present Danet-FCN3, a deep neural network specifically designed to classify seismic images at pixel level resolution. We have recently demonstrated that our deep learning models can distinguish among different rock layers helping the expert to interpret new seismic images. The dataset preparation processes the raw post-stacked data and the interpretation labels to produce training, validation and testing sets.